img×(-a)↓a↓{+/,g×⍵}⌺s⊢(⍪⍨⍪⊢)(,⍨,⊢)img

is it possible to improve this computation? it does stencil on a torus by padding. I tried the stencell mentioned in dyalog blog but that gives a WS FULL on 64 64 array.

What is s?

I guess it's convolution size

But I don't expect that to give WS FULL, as (⍪⍨⍪⊢)(,⍨,⊢)img should give a (64×3)(64×3) array which is not that large

s is 1+2×⌊2÷⍨3 5, where img is a 2D array

Btw 1+2×⌊2÷⍨x y is simply x y if x and y are odd

An obvious improvement would be changing (⍪⍨⍪⊢)(,⍨,⊢) to pad just what you need

x y is even.

let me try if trim the padded array would improve the performance

yes, img×(-h)↓(h←a÷2)↓{+/,g×⍵}⌺s⊢(2⊖⍪⍨)(2⌽,⍨)img gives 1 third improvement on performance

thank you for the suggestion

oh, I made a typo, s is 1+2×⌊2÷⍨⍴img

the roughly same size of img rounded to have odd edge sizes

Job opportunities: Do you like watching APL videos? Check out the two new listings, one for volunteer work and one paid internship!